Learning boosting: enhancing predictive modelling in blended learning environments

Accurate prediction of student learning outcomes is critical for early intervention and instructional decision-making in blended learning environments. This study proposes learning boosting, a structure-enhanced predictive framework integrating community-aware Louvain clustering with a gradient boosting classification. Student activity graphs are clustered to detect latent behavioural communities, and the resulting structural labels are embedded as features for final prediction. Experiments on real-world data from a blended learning course with 102 students evaluate the method under multiple classification granularities, data modalities, and clustering strategies. Results show that learning boosting consistently outperforms 11 baseline models, achieving an F1-score of 0.892, AUC of 0.883, and recall of 0.903 in the three-class task. Ablation studies confirm the complementary benefits of structural feature extraction and clustering. The findings demonstrate that combining graph-based structural modelling with boosting classifiers offers a robust and interpretable approach to learning analytics, especially in sparse and multimodal conditions.